TCP Congestion Control Performance on a Highway in a Live LTE Network

This paper investigates the behavior of several popular congestion control algorithms in LTE networks. Since TCP does not differentiate between different types of connections, the same loss-based congestion control algorithms are usually used with Ethernet, WiFi, and LTE. However, because packet losses are often concealed in LTE networks, the performance of such algorithms can be very poor, especially in mobile scenarios typical for LTE. We use measurements to analyze the behavior of these algorithms in a static case as well as a mobile scenario where the user equipment moves away from the eNodeB on a highway. In our measurements, we found that loss-based algorithms like Reno and CUBIC had trouble adjusting to quick changes of capacity due to their unawareness of packet losses, which also led to high delays. The combined delay- and loss-based approach of Compound TCP also performed poorly, as only the loss-based component was active most of the time. BBR (Bottleneck bandwidth and round-trip propagation time), a congestion control algorithm that adjusts its window based on changes in round-trip-times, was able to react to a changing capacity very quickly and therefore did not lead to additional delays like the other algorithms analyzed in this work.